Fix for Missing Dependencies (ModuleNotFoundError) when running Vevo2 AR
If you are trying to run the Vevo2 inference script (python -m models.svc.vevo2.infer_vevo2_ar) and keep getting hit with ModuleNotFoundError loop, the official requirements.txt is currently missing several key packages.
Additionally, the code relies on an older version of transformers, so you will get a ValueError: LlamaConfig accepts only keyword arguments, but found 5 positional args if you use the latest version.
To save you from playing dependency whack-a-mole, here is the consolidated fix.
Run this single command to install all the missing packages:
pip install accelerate torchvision torchaudio einops json5 ruamel.yaml six ipython pyworld librosa pydub scipy soundfile omegaconf tqdm gdown tensorboard
Run this command to fix the LlamaConfig / Transformers error:
pip install "transformers<4.40.0"
(Note: The ruamel.yaml package uses a dot, even though the error trace calls for ruamel_yaml with an underscore).
Hope this helps anyone else trying to get the AR + FM pipeline running locally
Hi Berserq, thanks for sharing! I am currently trying to run the inference, but I've run into a persistent issue when loading the tokenizer.
The Error:
When the pipeline attempts to initialize the tokenizer, it fails with the following exception:
fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)
Exception: data did not match any variant of untagged enum ModelWrapper at line 843742 column 3
Environment Details:
tokenizers == 0.15.2
transformers==4.39.0
Python: 3.10
CUDA Version: 12.8
What tokenizer and transformer version you are using? Thanks in advance for your help!